Reproduce “MMPerspective: Do MLLMs Understand Perspective? A Comprehensive Benchmark for Perspective Perception, Reasoning, and Robustness” on OpenResearch

About this paper

Yolo Y. Tang, Pinxin Liu, Zhangyun Tan, Mingqian Feng, Rui Mao, Chao Huang, Jing Bi, Yunzhong Xiao and 6 more

A new benchmark, MMPerspective, systematically evaluates how multimodal large language models process visual perspective. It demonstrates that while larger MLLMs show improved overall performance, they face considerable challenges in geometric reasoning and maintaining robust spatial interpretations when presented with visual alterations.

Key insights

• Larger MLLMs generally perform better on perspective tasks, but there is a significant performance drop between perception and compositional reasoning tasks, indicating difficulty with higher-level geometric inference.
• MLLMs exhibit poor robustness, often failing to maintain consistent predictions under simple geometric transformations like horizontal flipping, suggesting limited internalization of spatial priors and geometric constraints.
• Chain-of-Thought (CoT) prompting consistently improves performance and robustness across most tasks, while architectural choices introduce systematic biases in how models interpret perspective.

Summary by alphaXiv.